Tennis matches

Courtside deep v3. New deep reports use the ranking baseline plus the last 10 captured singles results, comparable serve/return statistics, and age-discounted head-to-head evidence where available. Combined non-ranking adjustments are capped at ±0.45 log-odds. The model runs 10,000 seeded sensitivity simulations across a declared ±12–20 percentage-point range; these are not calibrated confidence intervals. Injury, crowd and handedness context do not receive numerical adjustments. Original inputs and model version are saved with each prediction. Retirements remain in the displayed history but are excluded from numerical form. Incomplete history removes the form adjustment.

Legacy Courtside rank + form v1. Earlier forecasts remain visible and continue to settle. This experimental statistical baseline with fixed, hand-set weights; it is not trained or calibrated. It uses dated rankings and up to five completed matches in the prior 30 days. It does not assess injuries, surface, tactics or opponent quality.

New rank/form v2 fallback forecasts use the same captured ten-match sample. Earlier deep v2 and rank/form v1 predictions retain their original five-match, 30-day policy and recorded inputs.

The rank/form formula is: z = 0.7 × ln(rank B / rank A) + 0.6 × (form A − form B). Form = (wins + 2) / (matches + 4). P(A) = 1 / (1 + exp(−z)), limited to 10–90%. P(B) = 1 − P(A). Rankings older than 35 days cause an abstention; an unavailable form feed removes the form term.

Only the first eligible forecast per match is scored. Start status is checked again after collecting inputs. Same-day matches without a verified future start time are excluded from new forecasts. Existing searches are not backfilled with invented predictions. Slider changes do not modify the forecast.

Paper strategy: buy 10 Yes contracts on the eligible side with the highest positive expected profit after fees; otherwise pass. A fresh exact-match ask must show at least 10 contracts. The predicted winner and the value selection can differ. These are simulated entries, not confirmed fills or actual bets. V2 and V3 additionally require moderate evidence and positive EV at the conservative end of its sensitivity range. Limited evidence results in a pass. V1 entries retain their original rules.

Kalshi fees: single-price direct-member taker estimate, 0.07 × multiplier × quantity × price × (1 − price), with current model and balance rounding. The effective series/event fee multiplier is saved at entry. Actual multi-fill execution and account rounding may differ. Fee schedule · Rounding rules.

Paper P&L = contracts × final Kalshi payout − saved entry debit (including fees). ROI uses total settled net profit divided by total settled entry debit. Open positions are excluded from settled ROI. Walkovers and scalar settlements do not count toward accuracy or Brier score; their actual exchange payouts still count toward paper P&L. Played retirement wins can be scored; retired matches are excluded from the form input sample.

Brier score is the mean squared error of the saved probabilities: 0 is perfect, lower is better; a constant 50% forecast scores 0.25. Accuracy excludes exact 50% forecasts. Results describe only the matches you opened and recorded, not all tennis matches.